Research On Low-light Image Enhancement Method Based On Generative Adversarial Network

With the development of electronic shooting and the popularization of smart devices,smart shooting is widely used in smart cities and public transportation construction.However,due to the impact of low light conditions such as low light in reality,nighttime,occlusion,etc.,and the hardware limitations of low-precision image imaging equipment,the captured images often have problems with low brightness,serious noise,loss of detailed information,or blurring.As a result,the content of the picture is lost,hindering further understanding and analysis.Therefore,studying low-light image enhancement has important practical application value.In order to improve the perceived quality of the enhanced image,this article attempts to introduce a brightness attention mechanism and explore the construction of a lightweight image enhancement model.The main contents are summarized as follows:1.Use the attention mechanism to guide the enhancement of low-light images.This thesis proposes a low-illumination image enhancement method based on the brightness attention mechanism to generate an adversarial network.This method uses brightness attention to predict the light distribution in the low-light images,thereby guiding the enhancement network Adaptive enhancement of different brightness areas in the image.At the same time,based on the residual connection,a deep enhancement network is designed to enhance the enhancement process of modeling low-light pictures.In synthetic data sets,real data sets(DPED,LOL)verify algorithm performance,and compare with traditional image enhancement methods(histogram equalization,reflected illumination estimation),deep learning methods(DSLR),experiments show that our proposed network is enhanced The pictures have higher signal-to-noise ratio and structural similarity,and the overall perceived quality is relatively good,indicating the effectiveness of our method in low-light image enhancement.2.Aiming at the problem of large amount of calculation and many parameters in the existing low-illumination image enhancement network,this thesis proposes a low-illumination image enhancement method based on improved depth separable convolution to generate an adversarial network,which introduces depth separable convolution To improve the network structure for image enhancement tasks,two modules,improved depth separable convolution(IN-DepthwiseConv)and improved inverse residual depth separable convolution(IN-Bottleneck),are designed to reduce computational complexity and model parameter.At the same time,three models DwDG,DwG and DeeperDwG were constructed to verify the algorithm performance under the same data set and environment.The results show that our method can effectively reduce the amount of model parameters and maintain a better image enhancement effect.